eeg based patient monitoring system for mental alertness
TRANSCRIPT
EEG Based Patient Monitoring System for
Mental Alertness Using Adaptive Neuro-Fuzzy
Approach
Dilshad Begum Dept of ISE, Ghousia College of Engineering, Ramanagar, Bangalore, India
K. M. Ravikumar, Special Officer, VTU-Regional Office, Mysore, India
James. Mathew and Sanjeev Kubakaddi ITIE Knowledge Solutions, Bangalore, India
Rajeev Yadav Genia Photonics Inc, Laval, Canada
Abstract—Recent electrophysiological studies support
command-specific changes in the electroencephalography
(EEG) that have promoted their intensive application in the
noninvasive brain computer interfaces (BCI). However,
EEG is plagued by a variety of interferences and noises,
thereby demanding better accuracy and stability for its
application in the neuroprosthetic devices. Here we
investigate wavelets and adaptive neuro-fuzzy classification
algorithms to enhance the classification accuracy of
cognitive tasks. Using a standard cognitive EEG dataset, we
demonstrate improved performance in the classification
accuracy with the proposed system.
Index Terms— BCI, EEG, wavelets, adaptive neuro fuzzy
interface system (ANFIS)
I. INTRODUCTION
A patient recovering following a stroke, severe brain
damage or spinal injury may suffer from unresponsive
wakefulness syndrome (or vegetative state) [1]. Many of
these patients remain unresponsive for prolonged period
of times while some with limited cognitive capabilities.
Rehabilitation of such patients is extremely challenging
due to the difficulty in interpreting their cognitive
abilities. The brain-computer interfaces (BCIs) which
establishes a direct link between a human brain and a
computer discovers the potential of computers to enhance
physical and mental abilities; and often refers to the
interaction of human brain with external devices [2]-[4].
The human brain is a signal generator, sending out
millions of signals which can be measured and analyzed.
Modern biomedical research has provided the method for
extracting these brain signals in a convincing way to
distinguish their characteristics from one another. It was
Manuscript received August 21, 2013; accepted February 24, 2014.
soon discovered that these signals could be used to
communicate with an external peripheral device, through
proper machine learning techniques. Now, BCIs are often
directed at assisting, augmenting, or repairing human
cognitive or sensory-motor functions [5], [6].
Every time we think, move, feel or perform an activity,
the neurons in our brain are at work. This work is carried
out by small electric signals that transmit signals from
neuron to neuron. BCI technology monitors this electrical
activity, usually via the method of
electroencephalography (EEG). EEG characteristics are
then detected via special signal processing algorithms,
and the user is able to classify and generate
communication messages. BCI enables physically
disabled persons to perform many activities, thus
improving their quality of life and allowing them to be
independent [2]-[4], [7]-[9]. A general BCI system uses a
set of EEG samples for training and random samples for
testing the trained classifier. However, the classification
accuracy of the system performance is a great challenge.
Identifying the best discriminating features in the EEG
which is plagued by a variety of noise; in itself is a
challenge then selecting a classification technique
capable to distinguish a variety of time-varying patterns
is another challenge. Widespread research and
experiments have been conducted on algorithms to
classify brain signals according to its characteristic
features.
EEG patterns have been exploited widely in prosthetic
control [10]; for example, they have been used to move
the cursor on a computer display [11], to control hands-
free or virtual wheelchairs [12], [13] and to move robotic
limbs [13]-[17]. Different feature extraction techniques,
such as autoregressive coefficients [18]-[22]; component
analyses; spectral, spatial and correlation features[5], [6],
[21], [23]-[26]; transform features (short-time (windowed)
Journal of Medical and Bioengineering Vol. 4, No. 1, February 2015
©2015 Engineering and Technology Publishingdoi: 10.12720/jomb.4.1.59-66
59
Fourier wavelet, and affine transform) [26]; and time-
frequency-ambiguity features [19] have been
experimented on using EEG signals with standard
clustering and classification methods, including: support
vector machines , K-nearest neighbor, linear discriminant
analysis, Bayesian Classifier, Hidden Markov Model,
neural networks , fuzzy classifiers, genetic algorithm, etc
[27]-[36].
Adaptive Neuro-Fuzzy system is used in [18], [31],
[37]-[40] to detect P-300 rhythm in EEG, which are
obtained from visual stimuli followed by a motor
response from the subject. The paper explored wavelet
transform to extract features together with ICA algorithm.
41] proposed an intelligent system for ECG
classification based on ANFIS theory utilizing ICA
feature extraction. ANFIS algorithm is utilized by
37] to distinguish between imaginary left
and right hand motion using wavelet features, in which
the impact of cluster radius and number of fuzzy rules in
determining classification accuracy is investigated.
42] used five ANFIS classifiers to
distinguish between five different mental tasks and
compared the results with different neural network
architecture like MLP and Hierarchical Hybrid Model. A
43] compared the performance of
different neural network algorithms in classifying EEG
for five mental tasks with respect to wavelet transform
44] explored auto regression
features with neural networks for the discrimination of
45], [46] described a system
for on-line controlling the hand movement in a virtual
reality environment using Real-Time Recurrent
Probabilistic Neural Network.
EEG is considered the best suited potential non-
invasive interface, mainly due to its fine temporal
resolution, ease of use, portability and low set-up cost.
The project focuses on extracting characteristic features
of EEG signals to control an external device, say alarm
system. The research investigates EEG preprocessing,
wavelet feature extraction and Neuro- Fuzzy
classification to control an alarm system, alerting about
one’s mental state.
II.
METHOD
A.
Data Collection
In this research, a standard dataset from Purdue
University has been used. EEG Recording was performed
with a bank of Grass 7P511 amplifiers whose bandpass
analogue filters were set at 0.1 to 100 Hz. Signals are
taken using six electrodes in the central, parietal and
occipital regions - C3 C4 P3 P4 O3 O4 with respect to
10-20 electrode system. EOG signal is recorded
separately and the signals are claimed to be free from
artifacts. Data is collected for two cognitive tasks –
mental letter composing and geometric figure rotation. A
single trial recorded consists of 2500 samples at a
sampling frequency 250Hz. Ten trials of each task is
considered for this study.
Figure 1. Algorithm flowchart for the proposed BCI system.
B. Preprocessing and Feature Extraction
Some pre-processing techniques are to be applied to
the raw EEG signals to reduce error as well as
computational complexity. While recording, EOG and
line frequency artifacts are already removed in the dataset
[47].
Normalization: All the signals have to be
normalized to the range of [-1 1] to make the
calculation simpler.
Filtering: The signals are then fed to an equiripple
bandpass filter set to the range 0-40 Hz [6]. We
are interested only in delta, theta, alpha, beta and
lower gamma band signals, which have actionable
information.
Wavelet feature extraction: Wavelet features are
widely accepted and utilized due to high
discriminating properties in the active band (delta,
theta, alpha, beta, and gamma) for a particular
EEG condition. Wavelet decomposition gives
coefficients in these band areas, which clearly
signifies the brain signal. We considered wavelet
decomposed band coefficients to represent each
signal and to reduce the dimension of feature
matrix statistical parameters that has been selected.
Wavelets divide the signal into different frequency
components in a multi-resolution manner. As the EEG
signals are non-stationary, Fourier Transform may not be
a good choice. Since we couldn’t fix a permanent
window size for EEG patterns a time-frequency
representation is adopted to distinguish the subsequent
frequency components [48].
In wavelet transform, the signal is fed to filter bank
structure, a series of low pass and high pass filters. To
analyze the signal at different scales, different filter cutoff
frequencies are chosen for the sub filters. The resolution
of the signal, which is a measure of the amount of detail
information in the signal, is changed by the filtering
operations, and the scale is changed by up sampling and
down sampling (subsampling) operations.
Journal of Medical and Bioengineering Vol. 4, No. 1, February 2015
©2015 Engineering and Technology Publishing
Darvishi et al [
Barbosa et al [
Nazmy et al [
study by Khare et al [
features. Charles et al [
mental tasks. Ahamadi et al [
60
Figure 2. Three level decomposition tree of DWT [49].
The signal can be represented using a Mother wavelet -
( )x and its child wavelets, which are scaled and time
shifted versions of the mother wavelet [48].
2( , ) ( ) 2 2
s
s
s l x x l
(1)
The mother function ( )x is modified with respect to
variables s and l to generate a series of wavelet family.
The scale index s indicates the wavelet's width, and the
location index l gives its position. The mother functions
are rescaled, or dilated by powers of two, and translated
by integers. Thus, the mother wavelet basis functions
give a clear picture of the signal characteristics.
The wavelet transform is the convolution between the
two functions x and , which can be done in Fourier
space using the Fast Fourier Transform (FFT) [48]. In the
Fourier domain, the wavelet transform is simply,
k
1i n t
k
0
( ) ( ) *(s )eN
n
k
W s x k
(2)
where the ^ indicates the Fourier transform (FT), and the
Fourier transform of the time series is given by:
12 /
0
1 Nikn N
k n
n
x x eN
(3)
To use this formula, the FT of the wavelet function
should be known analytically. The wavelets are
normalized as,
1
22( ) ( )k o k
ss s
t
(4)
After preprocessing, the signals are subjected to
wavelet decomposition and the coefficient features are
extracted. The signal is decomposed to 6th level to
extract the bands of interest using Daubechies wavelets
(db14). Since the coefficient matrix is very large, we
apply some statistical data reduction techniques. The final
feature set will be a set of minimum of coefficients,
maximum of coefficients, standard deviation, mean and
power of the coefficient matrix (for each band-alpha, beta,
gamma, delta , theta). Thus for a single electrode signal, a
set of 25 features are extracted. Features are extracted for
8 trials each of two tasks- geometric figure rotation and
letter composing.
C. Classification
The algorithm is trained for a set of data (EEG signals)
which is been already taken from different subjects for
two particular tasks. Since the task is clear identification
and grouping of signal pattern, classification algorithms
are preferred to regression methods. Out of the different
classification algorithms like LDA, KNN, SVM, LVQ,
NN, Neuro-Fuzzy etc, we select classifiers with higher
adaptability and precision; i.e. ANFIS.
1) Neural fuzzy system
The term Neuro-Fuzzy refers to combinations of
artificial neural network and fuzzy logic (Fig 3). Neural
fuzzy systems provide a kind of automatic tuning method
to fuzzy systems by the use of neural networks, but
without altering their functionality. Neural networks can
be used for the membership function elicitation and
mapping between fuzzy sets that are utilized as fuzzy
rules.
Figure 3. Neuro-fuzzy system [50].
During the training process, a neural network adjusts
its weights in order to minimize the mean square error
between the output of the network and the desired output.
In Neuro-Fuzzy systems, the weights of the neural
network represent the parameters of the fuzzification
function, fuzzy word membership function, fuzzy rule
confidences and defuzzification function respectively
[50]. Thus the training of neural network results in
automatically adjusting the parameters of a fuzzy system
and finding their optimal values.
2) Adaptive neuro-fuzzy inference system (ANFIS)
Adaptive Neuro-Fuzzy Inference Systems are a class
of adaptive networks that are functionally equivalent to
fuzzy inference systems, incorporating the learning
capabilities of neural networks and the robustness of
fuzzy.
Fuzzy interface system maps input characteristics to
input membership functions and again this to a set of
rules. Rules could be set by the user’s interpretation of
the model characteristics. The rules are mapped to output
membership functions and then to an output value.
Neuro-adaptive learning techniques provide a method for
the fuzzy modelling procedure to learn about a data set. It
calculates the best membership function to map the given
data using neural network structure. The membership
function parameters changes throughout the learning
process, according to a gradient vector, which describes
the performance index of the fuzzy modelling [51].
Sugeno-Tsukamoto fuzzy model is a widely accepted
neuro-fuzzy system. It is a universal approximator. The
neuro-fuzzy system in this research follows a hybrid
learning algorithm, Least Mean Square Algorithm for
forward pass and Scaled Conjugate Gradient Method for
backward pass. For a fuzzy inference system with two
inputs x and y and one output z, a first-order Sugeno
fuzzy model has rules as the following:
Journal of Medical and Bioengineering Vol. 4, No. 1, February 2015
©2015 Engineering and Technology Publishing 61
Input nodes (Layer 1): All nodes are adaptive nodes.
The outputs are fuzzy membership grade of the inputs,
given by
1 ( ), 1,2i iO A x for i (5)
1
2( ), 3,4i iO B y for i (6)
where x and y are the inputs to the node I, A is a
linguistic label (small, large) and ( )iA x and
2 (y)iB can adopt any fuzzy membership function.
For a Gaussian (bell)-shaped function,
2
1( )
1 ( / )iA x
x ci ai b
(7)
where (ai, bi and ci) are the parameters of the
membership function. Parameters are referred to as
premise parameters.
Rule nodes (Layer 2): Rule layer is a fixed node
labelled M whose output is the product of all the
incoming signals. The outputs of this layer are the firing
strengths of the rule, represented as:
2 ( ) (y), i 1,2i i i iO w A x B (8)
Average nodes (Layer 3): In normalization layer also
nodes are fixed nodes, labelled N, which gives
normalized firing strengths
3
1 2/ (w w ), i 1,2i i iO w w (9)
Consequent nodes (Layer 4): It is defuzzification layer
in which all nodes are adaptive nodes. The output of each
node is simply the product of the normalized firing
strength and a first order polynomial.
4
i i i(p x q x r ), i 1,2i i i iO w f w (10)
Output nodes (Layer 5): This is summation neuron, a
fixed node which computes the overall output as the
summation of all incoming signals.
5
1 21/ ( )i i i i ii
O w f w f w w
(11)
3) Scaled conjugate gradient method (SCG)
SCG is a supervised learning algorithm for
feedforward neural networks, and is a member of the
class of conjugate gradient methods.
Figure 4. ANFIS structure [7].
Let expE
pT
be a vector from the space
.degT T , where N is the sum of the number of
weights and of the number of biases of the network. Let
be E be the error function we want to minimize [52].
Contrary to other conjugate gradient methods (CGMs),
SCG differs in two points:
Each iteration k of a CGM computes iw , where
NR is a new conjugate direction, and
1 .k k k kw w p is the size of the step in this
direction. Actually kp is a function of k , the
Hessian matrix of the error function, namely the
matrix of the second derivatives. In contrast to
other CGMs which avoid the complex
computation of the Hessian and approximate
k with a time-consuming line search procedure,
SCG makes the following simple approximation
of the term "( )kE w , a key component of the
computation of :k ks .
As the Hessian is not always positive definite,
which prevents the algorithm from achieving good
performance, SCG uses a scalar k which is
supposed to regulate the indefiniteness of the
Hessian. This is a kind of Levenberg-Marquardt
method, and is done by setting:
"( ).
'( . ) '( ),0 1
k k k k k
k k k kk
k
s E w p p
E w p E w
(12)
and adjusting k at each iteration. This is the main
contribution of SCG to both fields of neural learning and
optimization theory.SCG has been shown to be
considerably faster than standard back propagation and
than other CGMs [52].
The research adopts Neuro-Fuzzy classifier for
classifying the feature matrix into two groups. In this we
use the k-means algorithm to initialize the fuzzy rules.
For that reason, we should give the number of clusters for
each class. Here all the input features are grouped to one
cluster for each class. A Gaussian membership function is
used for fuzzy set descriptions, because of its simple
derivative expressions
The Scale gradient classifier based is on Jang’s neuro-
fuzzy classifier [51]. The differences are about the rule
weights and parameter optimization. The rule weights are
adapted by the number of rule samples. The scaled
conjugate gradient (SCG) algorithm is used to determine
the optimum values of nonlinear parameters. The SCG is
faster than the steepest descent and some second order
derivative based methods. Also, it is suitable for large
scale problems [53], [54].
4) K-means clustering:
Journal of Medical and Bioengineering Vol. 4, No. 1, February 2015
©2015 Engineering and Technology Publishing 62
The K-means algorithm divides the data into k
mutually exclusive clusters with actual observations. The
clusters are defined by its member objects and centroid.
The sum of distances from all objects in that cluster is
minimized at the centroid. The objects are moved towards
cluster centres using distance measurements like
Euclidean, Mahalanobis, city block etc. and is iterated till
the set of clusters become well- separable.
D. Testing Phase
Once the training is done, a fis structure is created with
two classes EEG signals of random task are given to the
system for testing The signals are pre-processed and
wavelet features are extracted. This is tested with the fis
structure and the neuro-fuzzy system returns the degree
of closeness to either of the classes. This degree of
closeness is rounded and is used to generate activation
signals to control external device. If the test signal is
grouped to any thought process, the alarm is set and alerts
about the mental activeness of the subject.
E. External Device Control
The objective of the project is to control an alarm and
led with respect to the classifier output. We use the
Arduino UNO board to set up the external device.
The Arduino UNO board with ATmega328
microcontroller is used to facilitate external device
control. In the board pin6 is connected to LED and pins 7,
9 and 11 are connected to 8ohm speakers. A simple
program is loaded to the controller to activate the output
pins when a digital 1 is coming through the serial port.
Speaker ports are programmed with suitable delay to
sound the output as an alarm. The system is tested
successfully for random test signals.
TABLE I. ACTIVATION SIGNAL FOR OUTPUT CONTROL
Index FIS Structure Activation Signal Device Control
1 Letter composing 0 No action
2 Figure Rotation 1 LED is ON
Alarm is ON
III. RESULTS AND DISCUSSION
The project investigates the relation between EEG
features and Fuzzy rules connecting them. Four different
clustering techniques- K means, grid partitioning,
subtractive clustering and Fuzzy C means are
experimented with respect to two backpropogation
network training algorithms, Levenberg-Marquardt and
Scaled gradient conjugate.
Figure 5. Training error for the SGC.
Figure 6. SGC membership function before and after training.
Figure 7. Fuzzy rules for the SGC network
Journal of Medical and Bioengineering Vol. 4, No. 1, February 2015
©2015 Engineering and Technology Publishing 63
(a): Grid Partitioning (b): FCM Clustering (c): Subtractive Clustering (d): Proposed SGC- k-means Clustering
Figure 8. Compares the membership Functions, Surface mesh and the network models of LM algorithm with (a) Grid Partitioning (b) Fuzzy C means (c) Subtractive Clustering(d) Proposed SGC algorithm with K means for EEG pattern classification.
The system is trained for 2 tasks over 500 iterations.
1) SCG-K means clustering:
For SGC algorithm, the training and testing
recognition rate got converged to 100% and 92.5%, the
mean square error got reduced to 0.0353 as shown in Fig.
5.
The membership functions that map the input to the
rule base are user defined Gaussian functions and got
reconfigured with the neural training epochs as shown in
Fig 6. Initially the membership function is randomly
defined, after setting the rules and fuzzification, the
membership function got reshaped to accommodate range
all range of feature values through the selected input
nodes. For each input set the membership function
defines two clusters for two classes, utilizing K-mean
clustering technique.
Fuzzy sets AND rules for all the input clusters and
classifies into two groups. These rules are put into the
network as weights and after every training loop, the
input functions got adjusted since the fuzzy weights are
set to be strict and not subjected to any variations. The
fuzzification process is detailed in Fig. 7.
2) Levenberg Marquardt- grid partitioning, fuzzy C
means, subtractive clustering:
The Levenberg-Marquardt algorithm approaches the
second-order training speed without calculating Hessian
matrix. The Levenberg-Marquardt algorithm
approximates Hessian matrix in the following Newton-
like update:
1
1 [ ]T T
k kX X J J I J e
(13)
where J is the Jacobian matrix that contains first
derivatives of the network errors with respect to the
weights and biases, and e is a vector of network errors.
Using a standard backpropagation technique Jacobian
matrix can be calculated, that is much simpler compared
to computing the Hessian matrix. When μ=0, this is
Newton’s method, for large μ, this becomes a gradient
descent with small step size. After each step μ is
decreased to reach near error minimum and is increased
only when a step would increase the performance
function. Thus the performance function got reduced in
each iteration step.
In subtractive clustering, each data point is assumed to
be a potential cluster center and a measure of the
likelihood that each data point is calculated based on the
density of surrounding data points. The data point with
the highest potential is selected to be the first cluster
center. Then all data points in the vicinity of the first
cluster center (as determined by radii) are removed so as
to determine the next data cluster and its center location.
This process is iterated until all of the data is within radii
of a cluster center [24].
In FCM each data point is grouped to a cluster to some
degree which is specified by a membership grade. The
algorithm starts with an initial guess for the cluster
centers, which is most likely incorrect. The cluster
centers and the membership grades for each data point are
updated iteratively till the cluster centers to the right
location within a data set. This iteration is subjected to
minimize an objective function which provides the
distance from any given data point to a cluster center that
is weighted by the membership grade of that data point.
TABLE II. COMPARISON OF TRAINING ERROR
LM-
Grid
LM-
Subtractive
LM-FCM SGC-K
means
Error
(500 iterations)
0.3383 0.4540 0.4525 0.0353
From Fig. 8, we can conclude that for FCM and
subtractive clustering, the network model is less complex
than grid partitioning. K means clusters follow straight
forward approach in getting the classifier output to either
of the two groups. The membership functions got
adjusted to the fuzzy rules during training and FCM and
subtractive methods form clusters for each set of input
features. While in K means, all the input features got
clustered into two classes, regardless the number of
features provided and the user defined Gaussian
membership functions got adjusted to accommodate the
varying data points. The surface mesh energy output is
more optimized in SGC method. The network gets
overloaded when the input features are separately fed to
the LM algorithm and it takes a long time to train the
fuzzy network, but SGM is observed to avoid that time
lag as it approaches the minimum error point through a
Journal of Medical and Bioengineering Vol. 4, No. 1, February 2015
©2015 Engineering and Technology Publishing 64
faster and optimum path. The training error for 500
iterations got compared in Table II, and SGC approach
gives a comparatively low error rate.
After the classification system is tested with random
signals and the degree of closeness to specific classes are
measured. This degree of closeness is used to group the
signal into letter composing and figure rotation. For
letter composing the system remain in the same state but
for figure rotation an alarm signal is evoked.
IV.
In this paper, wavelet based feature extraction is
incorporated with Adaptive Neuro- Fuzzy Interface
System classifier and various clustering and training
algorithms got compared. The Neuro-Fuzzy classifier
provides the information about the link between input
features and the relationship with corresponding classes,
adopting data clustering logic to form well-separable
groups. The SGM and LM algorithms optimize the
network to reduce the classification error, and it is seen
that the SGM gives better classification results. The rule
based classifier is proved capable enough to distinguish
different mental activities which in turn activate the
external device. The project proves the possibility of
controlling external devices using mental thoughts in an
optimized way via Neuro –Fuzzy EEG processing.
ACKNOWLEDGMENT
The Authors would like to thank the University of
Purdue, USA, for providing the EEG data and ITIE Inc
for providing the necessary equipment’s to carry out the
experiments.
REFERENCES
[1]. T. Hinterberger, et al., "A device for the detection of cognitive
brain functions in completely paralyzed or unresponsive patients,"
IEEE Transactions on Biomedical Engineering, vol. 52, pp. 211-220, 2005.
[2]. C. S. Nam, G. Schalk, and M. M. Jackson, "Current trends in
brain–computer interface (bci) research and development," International Journal of Human-Computer Interaction, vol. 27, no.
1, pp. 1-4, 2010.
[3]. D. Mattia, et al., "Hybrid brain-computer interaction for functional
motor recovery after stroke," in Converging Clinical and
Engineering Research on Neurorehabilitation Biosystems &
Biorobotics, vol. 1, 2013, pp. 1275-1279. [4]. Q. Noirhomme, et al., "Detecting consciousness with a brain-
computer interface," in Converging Clinical and Engineering
Research on Neurorehabilitation Biosystems & Biorobotics, vol. 1, pp. 1261-1264, 2013.
[5]. R. Yadav, M. N. Swamy, and R. Agarwal, "Model-based seizure
detection for intracranial EEG recordings," IEEE Trans Biomed Eng, vol. 59, no. 5, pp. 1419-28, 2012.
[6]. R. Yadav, et al., "Morphology-based automatic seizure detector
for intracerebral EEG recordings," IEEE Trans Biomed Eng, vol. 59, no. 7, pp. 1871-1881, 2012.
[7]. I. U. E. D. Guler, "Adaptive neuro-fuzzy inference system for
classification of EEG signals using wavelet coefficients," Journal of Neuroscience Methods, vol. 148, pp. 113-121, 2005.
[8]. J.-H. Lim, et al., "Corrigendum: Classification of binary intentions
for individuals with impaired oculomotor function: ‘eyes-closed’
SSVEP-based brain–computer interface," Journal of Neural
Engineering, vol. 10, no. 3, pp. 039501, 2013. [9]. E. W. Sellers, "New horizons in brain-computer interface
research," Clin Neurophysiol, vol. 124, no. 1, pp. 2-4, 2013.
[10]. Z. H.-l. F. Zhen, "Prosthetic controlled system based on signal pattern recognition of electroencephalogram," International
Seminar on Future BioMedical Information Engineering, pp. 287-
290, 2008. [11]. M. Dong, et al., "Study on eeg-based mouse system by using
brain-computer interface, international conference on virtual
environments," Human-Computer Interfaces and Measurements Systems, pp. 236-239, 2009.
[12]. G. G. Gentiletti, et al., "Command of a simulated wheelchair on a
virtual environment using a brain-computer interface," IRBM, vol. 30, pp. 218-225, 2009.
[13]. A. Ferreira, et al., "Human-machine interface based on muscular
and brain signals applied to a robotic wheelchair," Journal of Physics, 2007.
[14]. E. W. Sellers, T. M. Vaughan, and J. R. Wolpaw, "A brain-
computer interface for long-term independent home use," Amyotroph Lateral Scler, vol. 11, no. 5, pp. 449-55, 2010.
[15]. M. N. R. Y. S. Murugappan, "Comparison of different wavelet
features from EEG signals for classifying human emotions," ISIEA, 2009, pp. 836-841.
[16]. C. S. Nam, J. Woo, and S. Bahn, "Severe motor disability affects
functional cortical integration in the context of brain-computer interface (BCI) use," Ergonomics, vol. 55, no. 5, pp. 581-91, 2012.
[17]. B. Z. Allison, et al., "Toward a hybrid brain-computer interface
based on imagined movement and visual attention," J Neural Eng, vol. 7, no. 2, pp. 26007, 2010
[18]. W. Y. Hsu, "EEG-based motor imagery classification using neuro-
fuzzy prediction and wavelet fractal features," J Neurosci Methods, vol. 189, no. 2, pp. 295-302, 2010.
[19]. W. Y. Hsu, "Embedded prediction in feature extraction:
application to single-trial EEG discrimination," Clin EEG Neurosci, vol. 44, no. 1, pp. 31-38, 2013.
[20]. Jin-Fu, L. Ming-Ai, et al., "Feature Extraction and Classification
of Mental EEG for Motor Imagery," in Fifth International Conference on Natural Computation, 2009.
[21]. R. Yadav, R. Agarwal, and M. N. Swamy, "A novel dual-stage
classifier for automatic detection of epileptic seizures," in Proc. Conf Proc IEEE Eng Med Biol Soc, 2008, pp. 911-914.
[22]. J. M. Ramírez-Cortes, et al., "P-300 Rhythm Detection Using ANFIS Algorithm and Wavelet Feature Extraction in EEG
signals," in Proc. World Congress on Engineering and Computer
Science, 2010. [23]. R. Yadav, R. Agarwal, and M. S. Swamy, "A novel morphology-
based classifier for automatic detection of epileptic seizures," in
Proc. Conf Proc IEEE Eng Med Biol Soc, 2010, pp. 5545-5548. [24]. R. Yadav, et al., "A novel unsupervised spike sorting algorithm
for intracranial EEG," in Proc. Conf Proc IEEE Eng Med Biol Soc,
2011, pp. 7545-7548. [25]. R. Yadav, M. N. Swamy, and R. Agarwal, "Rapid identification of
epileptogenic sites in the intracranial EEG," in Proc. Conf Proc
IEEE Eng Med Biol Soc, 2011, pp. 7553-7556. [26]. G. N. Garcia, T. E. and J.-M. Vesin, "Support vector eeg
classification in the fourier and time-frequency correlation
domains," in Proc. International IEEE EMBS Conference on Neural Engineering, 2003.
[27]. A. F. Cabrera, D. Farina, and K. Dremstrup, "Comparison of
feature selection and classification methods for a brain-computer interface driven by non-motor imagery," Med Biol Eng Comput,
vol. 48, no. 2, pp. 123-132, 2010.
[28]. M. A. Oskoei, J. Q. Gan, and H. Hu, "Adaptive schemes applied to online SVM for BCI data classification," in Conf Proc IEEE Eng
Med Biol Soc, 2009, pp. 2600-2603.
[29]. B. Hong, et al., "Adaptive active auditory brain computer interface," in Proc. Conf Proc IEEE Eng Med Biol Soc, 2009, pp.
4531-4534.
[30]. Q. Xu, et al., "Fuzzy support vector machine for classification of EEG signals using wavelet-based features," Med Eng Phys, vol. 31,
no. 7, pp. 858-65, 2009.
[31]. B. Rivet and A. Souloumiac, "Subspace estimation approach to P300 detection and application to brain-computer interface," in
Proc. Conf Proc IEEE Eng Med Biol Soc, 2007, pp. 5071-5074.
[32]. S. Raiesdana, et al., "Wavelet packet decomposition of a new filter -- based on underlying neural activity -- for ERP classification," in
Proc. Conf Proc IEEE Eng Med Biol Soc, 2007, pp. 1876-1879.
[33]. Y. Li and C. Guan, "A semi-supervised SVM learning algorithm for joint feature extraction and classification in brain computer
Journal of Medical and Bioengineering Vol. 4, No. 1, February 2015
©2015 Engineering and Technology Publishing 65
CONCLUSIONS
interfaces," in Proc. Conf Proc IEEE Eng Med Biol Soc,. vol. 1, 2006, pp. 2570-2573.
[34]. X. Liao, D. Yao, and C. Li, "Transductive SVM for reducing the
training effort in BCI," J Neural Eng, vol. 4, no. 3, pp. 246-54, 2007.
[35]. T. N. Lal, et al., "Support vector channel selection in BCI," IEEE
Trans Biomed Eng, vol. 51, no. 6, pp. 1003-1010, 2004. [36]. A. Vallabhaneni and B. He, "Motor imagery task classification for
brain computer interface applications using spatiotemporal
principle component analysis," Neurol Res, vol. 26, no. 3, pp. 282-7, 2004.
[37]. S. Darvishi and A. Al-Ani, "Brain-computer interface analysis
using continuous wavelet transform and adaptive neuro-fuzzy classifier," in Proc. Conf Proc IEEE Eng Med Biol Soc, 2007, pp.
3220-3223.
[38]. M. Xu, et al., "Channel selection based on phase measurement in P300-based brain-computer interface," PLoS One, vol. 8, no. 4, pp.
e60608, 2013.
[39]. B. Hong, et al., "N200-speller using motion-onset visual response," Clin Neurophysiol, vol. 120, no. 9, pp. 1658-66, 2009.
[40]. M. Salvaris and F. Sepulveda, "Visual modifications on the P300
speller BCI paradigm," J Neural Eng, vol. 6, no. 4, pp. 046011, 2009.
[41]. T. M. Nazmy, et al., "Adaptive neuro-fuzzy interface system for
classification of ECG signals," Journal of Theoretical and Applied Information Technology, 2009.
[42]. A. O. G. Barbosa et al., "Mental Tasks Classification for a
Noninvasive BCI Application," in Proc. 19th International Conference on Artificial Neural Networks, vol. 5769, 2009, pp.
495-504.
[43]. V. Khare, "Classification of five mental tasks based on two methods of neural network," International Journal of Computer
Science and Information Security, vol. 8, pp. 86-92, 2010.
[44]. W. A. Charles and Z. Sijercic, "Classification of EEG signals from four subjects during five mental tasks," in Proc. IEEE
Proceedings on Engineering Application in Neural Networks,
1997, pp. 407-414. [45]. A. Ahmadi, R. Jafari, and J. Hart, "Light-weight single trial EEG
signal processing algorithms: computational profiling for low power design," in Proc. Conf Proc IEEE Eng Med Biol Soc, 2011,
pp. 4426-4430.
[46]. M. Ahmadi and A. Erfanian, "An on-line BCI system for hand movement control using real-time recurrent probabilistic neural
network’," in Proc. 4th International IEEE Conference on Neural
Engineering, 2009. [47]. Z. A. Keirn, "Alternative modes of communication between man
and machine," Electrical Engineering Department, Purdue
University, 1998. [48]. C. Torrence and G. P. Compo, "A practical guide to wavelet
analysis," Bulletin of the American Meteorological Society, vol. 79,
no. 1, pp. 61-78, 1998. [49]. Wikipedia, Discrete Wavelet Transform, 2012.
[50]. T. A. Azhar, Adaptive Neuro-Fuzzy Systems, 2010, pp. 216
[51]. J. S. R. Jang, et al., "Neuro-fuzzy and soft computing: A computational approach to learning and machine intelligence,"
IEEE Transactions on Automatic Control, Prentice Hall, vol. 42,
no. 10, pp. 1482-1484, 1997. [52]. Introduction to SNNS: Scaled Conjugate Gradient, 2013 (Online).
[53]. B. Cetisli and A. Barkana, "Speeding up the scaled conjugate
gradient algorithm and its application in neuro-fuzzy classifier training," Soft Computing, vol. 14, no. 4, pp. 365–378, 2010.
[54]. H. Takahashi, T. Yoshikawa, and T. Furuhashi, "Application of
support vector machines to reliability-based automatic repeat
request for brain-computer interfaces," in Proc. Conf Proc IEEE Eng Med Biol Soc, 2009, pp. 6457-6460.
Dilshad Begum is a PhD candidate in Electrical and Electronics
Research Center Ghousia College of Engineering, Ramanagara Karnataka, under Visvesvaraya Technological University (VTU), India.
She received BE Degree in Computer Science and Engineering from
Gulbarga University Gulbarga in 2000 and MTech Degree from Visvesvaraya Technological University (VTU) in 2007 .Her current
area of interest is in EEG data Processing, Brain Computer Interface and
its Applications. ([email protected])
Ravi Kumar K. M Professor and Head of Information Science and Engineering Department, Ghousia collage of Engineering,
Ramanagaram. He has completed his M.Tech in the field of Biomedical
Instrumentation in the year 2002 and submitted his PhD thesis in 2009 (VTU, Belgaum). He is in the field of teaching from past 15 years and
he has published ten papers in the International Conference and four in
the International journal related to his research areas. He is member of various professional societies which include ISTE, MIE, SIS, and
BMESI. His field of interest includes Digital Signal Processing,
Human Computer Interface and Communication Systems. He is serving as Special Officer VTU, Belgaum ([email protected]).
Sanjeev Kubakaddi is successful entrepreneur and proprietor of itie Knowledge Solutions Bangalore, company focusing on Biomedical,
Embedded product development and technology training. He received
his Bachelors in 1995, Master of Technology (M. Tech) in 2012. He is also working towards his PhD from Jain University. He is a renowned
trainer in the areas of DSP, Image Processing and expert in Matlab®,
Simulink® and Texas Instruments DSP. Mr. Sanjeev Kubakaddi ([email protected]) is successful entrepreneur and proprietor of itie
Knowledge Solutions Bangalore, company focusing on Biomedical,
Embedded product development and technology training. He received his Bachelors in 1995, Master of Technology (M. Tech) in 2012. He is
also working towards his PhD from Jain University. He is a renowned
trainer in the areas of DSP, Image Processing and expert in Matlab®, Simulink® and Texas Instruments DSP ([email protected]).
James Mathew completed his Bachelors in Electrical & Electronics
Engineering from MG University, Kerala, India and Masters in
Communication Engineering & Signal Processing from University of Plymouth, UK. His research interests are Cognitive Neuroscience,
Neural Engineering, Brain Computer Interface, Image Processing,
Machine Learning and Pattern Recognition.
Rajeev Yadav received the B.Sc. degree in physics and the M.Sc. degree in electronics both from the D.D.U. Gorakhpur University,
Gorakhpur, India, in 1998 and 2000, respectively and Ph.D. in electrical
engineering from the Concordia University, Montreal, Canada in 2012. He is currently a Research Scientist at the Genia Photonics Inc., Laval,
Canada and a recipient of the prestigious NSERC Industrial R&D
fellowship of Canada. Prior to this, he worked as a Research Scientist at the Leap Medical Inc. Montreal, Canada. He is member of the IEEE
EMB, SP, CAS societies, CMBES, OSA and IBRO. He serves as
committee member for the CMBES and OSA, editorial member and reviewer for over 30 international signal processing and biomedical
engineering journals. His research interests include biophotonics,
biomedical signal processing and pattern recognition with emphasis on their application to neural signals for critical care and neuroscience
research, and neural prosthesis.
Journal of Medical and Bioengineering Vol. 4, No. 1, February 2015
©2015 Engineering and Technology Publishing 66